WO2023108437A1

WO2023108437A1 - Channel state information (csi) compression feedback method and apparatus

Info

Publication number: WO2023108437A1
Application number: PCT/CN2021/138032
Authority: WO
Inventors: 陈栋; 池连刚
Original assignee: 北京小米移动软件有限公司
Priority date: 2021-12-14
Filing date: 2021-12-14
Publication date: 2023-06-22
Also published as: EP4451575A1; CN118202584A

Abstract

Disclosed in the embodiments of the present application are a channel state information (CSI) compression feedback method and apparatus, which can be applied to various types of communication systems, for example, a 5th generation (5G) mobile communication system, a 5G new radio (NR) system, or other future new mobile communication systems. The method comprises: acquiring an estimated CSI image H of a network device, and generating a timing CSI image H_c according to the estimated CSI image H; compressing the timing CSI image H_c, so as to generate a feature code word; and sending the feature code word to the network device. A timing CSI image H_c corresponding to an estimated CSI image H is compressed by means of a terminal device, so as to generate a feature code word, and the timing CSI image is fed back to a network device by means of the feature code word. The channel resources occupied for feeding back a CSI image can be reduced, thereby saving on resources; and the precision of feeding back the estimated CSI image to a network device is improved.

Description

A method and device for channel state information CSI compression feedback

technical field

The present application relates to the field of communication technologies, and in particular to a method and device for compressing and feeding back channel state information (CSI).

Background technique

With the development of the fifth-generation wireless communication network, massive Multiple-Input Multiple-Output (mMIMO) has become a key technology. By configuring a large number of antennas, mMIMO not only greatly improves the channel capacity under limited spectrum resources, but also has a strong anti-interference capability. In order to make better use of mMIMO technology, the transmitter needs to obtain channel state information (Channel State Information, CSI). In the system, the terminal equipment estimates the CSI of the downlink channel, and then feeds the CSI back to the network equipment through a feedback link with a fixed bandwidth.

However, the multi-antenna property of mMIMO makes the overhead of CSI feedback huge, and there is still a lack of means to feed back CSI efficiently and accurately.

Contents of the invention

Embodiments of the present application provide a method and device for compressing and feeding back channel state information (CSI), which can be applied to various communication systems. For example: a fifth generation (5th generation, 5G) mobile communication system, a 5G new radio (new radio, NR) system, or other future new mobile communication systems. The terminal device compresses the time-series CSI image _Hc corresponding to the estimated CSI image H to generate a feature codeword, and feeds back the time-series CSI image to the network device through the feature codeword. The channel resource occupied by the feedback CSI image can be reduced, resources are saved, and the accuracy of the feedback CSI image is improved.

In the first aspect, the embodiment of the present application provides a method for channel state information CSI compression feedback, which is applied to a terminal device, and the method includes:

Obtain an estimated CSI image H of the network device, and generate a time-series CSI image _Hc according to the estimated CSI image H;

Compressing the time series CSI image _Hc to generate a feature codeword;

Send the feature codeword to the network device.

Optionally, the compressing the time-series CSI image _Hc to obtain a feature codeword includes:

The time-series CSI image _Hc is input into the self-information domain converter to generate a time-series self-information image _He , wherein the time-series CSI image _Hc and the time-series self-information image _He are both T in time dimension;

Input the time-series self-information image _He into a time-series feature coupling encoder for feature extraction to generate a structural feature matrix and a temporal correlation matrix;

The feature codeword is generated according to the structural feature matrix and the time correlation matrix.

Optionally, the inputting the time-series CSI image _Hc into a self-information domain converter to generate a time-series self-information image includes:

The time-series CSI image H _c is input into the three-dimensional convolutional feature extraction network to extract features to obtain the first time-series feature image F, wherein the convolution kernel specification of the three-dimensional convolutional network is f×t×n×n, and the f is the number of feature extractions, the t is the depth of convolution in the time dimension, and the n is the length and width of the convolution window;

generating a first index matrix M according to the time series CSI image _Hc ;

A time-series self-information image is obtained according to the first time-series feature image F and the first index matrix M.

Optionally, the generating the first index matrix M according to the time-series CSI image H _c includes:

Input the time-series CSI image _Hc into a self-information module to generate self-information of the area to be estimated in the time-series CSI image _Hc , and use it as a self-information image;

The self-information image is input into the index matrix module for mapping to obtain the first index matrix M.

Optionally, the step of inputting the time-series CSI image H _c into a self-information module to obtain self-information of the area to be estimated in the time-series CSI image H _c to obtain a self-information image includes:

Split the time-series CSI image H _c in time series to obtain split images H _c,i at each time point;

Divide the split image into multiple regions p _j to be estimated, and obtain self-information estimation values of the regions to be estimated

According to the estimated value from the self-information

Generated from the information image I _c,i .

Optionally, the index matrix module includes a mapping module network and a decision device, and the mapping of the self-information image input index matrix module to obtain the first index matrix M includes:

Inputting the self-information image into the mapping network to extract features to obtain a first information feature image D _c,i , wherein the mapping network is a two-dimensional convolutional neural network;

Inputting the first information feature image D _c,i into the decision device for binarization processing to obtain a second index matrix M _i ;

The second index matrix M _i is concatenated to obtain the first index matrix M.

Optionally, the mapping network includes a two-dimensional convolutional layer, a two-dimensional normalization layer, and an activation layer, and the inputting the self-information image into the mapping network to extract features includes:

inputting the self-information image into the two-dimensional convolutional layer to extract features to obtain a first feature image;

inputting the first feature image into the two-dimensional normalization layer to normalize the pixel values in the first feature image to obtain a second feature image;

Inputting the second feature image into an activation function layer for nonlinear mapping to obtain the first information feature image D _c,i .

Optionally, the splicing the second index matrix M _i to obtain the first index matrix M includes:

The second index matrix M _i is spliced in a time series order to obtain the first index matrix M.

Optionally, the acquiring a time-series self-information image according to the first time-series feature image F and the first index matrix M further includes:

multiplying the first time-series feature image F and the first index matrix M to obtain a second information feature image;

Inputting the second information feature image into a dimension restoration network to perform dimension restoration to generate the time-series self-information image _He .

Optionally, the temporal feature coupled encoder includes a one-dimensional space-time compression network and a coupled long short-term memory network LSTM.

Optionally, the time-series self-information image _He is input into a time-series feature coupling encoder for feature extraction to generate a structural feature matrix and a temporal correlation matrix, including:

The time sequence is input into the one-dimensional space-time compression network for one-dimensional space-time compression after dimension transformation from the information image _He to obtain the structural feature matrix, wherein the convolution kernel specification of the one-dimensional space-time compression network is S× 2N _c N _t ×m, the 2N _c N _t is the length of the convolution window, the m is the width of the convolution window, S is the target dimension, and the dimension of the structural feature matrix is T×S.

Optionally, inputting the time-series self-information image He _into a time-series feature coupling encoder for feature extraction to generate a structural feature matrix and a temporal correlation matrix, further comprising:

After performing dimension transformation on the time series from the information image _He , input coupling LSTM to extract features to obtain the temporal correlation matrix, wherein the dimension of the temporal correlation matrix is T×S;

The structural feature matrix and the temporal correlation feature matrix are coupled to generate the feature codeword.

Optionally, also include:

Input the training time-series CSI image _Hc into the self-information domain converter to obtain the training time-series self-information image _He ;

Input the training time-series self-information image He _into the time-series feature coupling encoder to obtain the training feature codeword.

Optionally, also include:

Send the training data to the network device, wherein the training data includes the training feature codeword, the time sequence length of the time-series self-information image _He , the dimension of the training feature codeword and the training time-series CSI image _Hc .

In the second aspect, the embodiment of the present application provides another method for channel state information CSI compression feedback, which is applied to a network device, and the method includes:

Receive the feature code word sent by the terminal device;

Restoring the feature codeword to obtain the restored time-series CSI image

Restore timing CSI images according to the

Get the restored estimated CSI image

Decompress the characteristic code word compressed by the terminal device through the network device to obtain the restored restored estimated CSI image

The channel resource occupied by the feedback CSI image can be reduced, resources are saved, and the accuracy of the feedback CSI image is improved.

Optionally, the restoring the feature codeword includes:

Input the feature codeword into the decoupling module for decoupling, so as to obtain the restored time series self-information image

Restore the time series from the information image

Input the restored convolutional neural network for restoration to obtain the restored time-series CSI image

Optionally, the decoupling module includes a one-dimensional spatio-temporal decompression network and a decoupling LSTM, and the decoupling of the input decoupling module of the feature codeword to obtain the restored time series self-information image includes:

Inputting the feature codeword into the one-dimensional space-time decompression network for decompression to obtain the restored structure feature matrix;

Inputting the feature codeword into the decoupling LSTM for decoupling to obtain the restored time correlation matrix;

Acquiring the restored time-series self-information image according to the restored structural feature matrix and the restored time correlation matrix

Optionally, the convolution kernel specification of the one-dimensional space-time decompression network is 2N _c N _t × M × m, the T is the number of rows of the restored temporal correlation matrix, and the 2N _c N _t is the Describes the number of columns of the restored time dependence matrix.

Optionally, the obtaining the restored time-series self-information image according to the restored structural feature matrix and the restored time correlation matrix includes:

Adding the restored structural feature matrix and the restored time correlation matrix point-to-point, and performing dimension transformation to obtain the restored time-series self-information image

Optionally, the restored convolutional neural network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and a seventh convolutional layer. Convolution layer, wherein, the convolution kernel specification of the first convolution layer and the fourth convolution layer is l ₁ ×t×n×n, the convolution of the second convolution layer and the fifth convolution layer The kernel specification is l ₂ ×t×n×n, the convolution kernel specification of the third convolution layer, the sixth convolution layer and the seventh convolution layer is 2×t×n×n, and the t is time The depth of the convolution in the dimension, the l ₁ , l ₂ and 2 are the number of extracted features, and the n is the length and width of the convolution window.

Optionally, the restoring the time series from the information image

include:

Restore the time series from the information image

Input the first convolutional layer for convolution to obtain a first restored feature map, input the first restored feature map to the second convolutional layer to obtain a second restored feature map, and input the second restored feature map The third convolutional layer is used to obtain a third restored feature map, and the third restored feature map and the restored time-series self-information image

sum to obtain a fourth reduced feature map;

Inputting the fourth restored feature map into the fourth convolutional layer to obtain a fifth restored feature map, inputting the fifth restored feature map into the fifth convolutional layer to obtain a sixth restored feature map, and converting the The sixth restored feature map is input into the sixth convolutional layer to obtain a seventh restored feature map, and the fourth restored feature map and the seventh restored feature map are added to obtain an eighth restored feature map;

Inputting the eighth restored feature map into the seventh convolutional layer for normalization to obtain the restored time-series CSI image

Optionally, also include:

Receiving the training data sent by the terminal device, the training data includes the training feature codeword, the time sequence length of the time-series self-information image _He , the dimension of the training feature codeword and the training time-series CSI image;

Obtaining a restored time-series CSI image according to the training feature codeword;

Perform training according to the restored time-series CSI images and the training time-series CSI images.

Optionally, also include:

Determine the number of structural units in the decoupled LSTM according to the time series length of the time series self-information image _He ;

The network parameters of the one-dimensional space-time decompression network are determined according to the dimensions of the training feature codewords.

Optionally, also include:

Carry out multiple rounds of training, the formulation of the learning rate in the training is expressed as:

Wherein, the γ is the current learning rate, the γ _max is the maximum learning rate, the γ _min is the minimum learning rate, the t is the current training round, and the T _w is the number of gradual learning, so The T' is the number of overall training cycles.

Optionally, also include:

Obtain recommended network parameters of the decoupling module and restored convolutional neural network, and update the decoupling module and restored convolutional neural network according to the recommended network parameters.

In the third aspect, the embodiment of this application provides a communication device, which has some or all functions of the terminal equipment in the method described in the first aspect above, for example, the functions of the communication device may have part or all of the functions in this application The functions in the embodiments may also have the functions of independently implementing any one of the embodiments in the present application. The functions described above may be implemented by hardware, or may be implemented by executing corresponding software on the hardware. The hardware or software includes one or more units or modules corresponding to the above functions.

In an implementation manner, the structure of the communication device may include a transceiver module and a processing module, and the processing module is configured to support the communication device to perform corresponding functions in the foregoing method. The transceiver module is used to support communication between the communication device and other equipment. The communication device may further include a storage module, which is used to be coupled with the transceiver module and the processing module, and stores necessary computer programs and data of the communication device.

As an example, the processing module may be a processor, the transceiver module may be a transceiver or a communication interface, and the storage module may be a memory.

In an implementation manner, the communication device includes:

An estimation module, configured to acquire an estimated CSI image H of the network device, and generate a time-series CSI image _Hc according to the estimated CSI image H;

A compression module, configured to compress the time series CSI image _Hc to generate a feature codeword;

A sending module, configured to send the feature codeword to a network device.

In the fourth aspect, the embodiment of the present application provides another communication device, which can realize some or all of the functions of the network equipment in the method example mentioned in the second aspect above, for example, the functions of the communication device can have some of the functions in this application Or the functions in all the embodiments may also have the function of implementing any one embodiment in the present application alone. The functions described above may be implemented by hardware, or may be implemented by executing corresponding software on the hardware. The hardware or software includes one or more units or modules corresponding to the above functions.

In an implementation manner, the structure of the communication device may include a transceiver module and a processing module, and the processing module is configured to support the communication device to perform corresponding functions in the foregoing method. The transceiver module is used to support communication between the communication device and other devices. The communication device may further include a storage module, which is used to be coupled with the transceiver module and the processing module, and stores necessary computer programs and data of the communication device.

In an implementation manner, the communication device includes:

The receiving module is used to receive the characteristic code word sent by the terminal equipment;

A restore module, configured to restore the feature codewords to obtain restored time-series CSI images

A channel acquisition module, configured to restore time series CSI images according to the

Get the restored estimated CSI image

In a fifth aspect, an embodiment of the present application provides a communication device, where the communication device includes a processor, and when the processor invokes a computer program in a memory, it executes the method described in the first aspect above.

In a sixth aspect, an embodiment of the present application provides a communication device, where the communication device includes a processor, and when the processor invokes a computer program in a memory, it executes the method described in the second aspect above.

In the seventh aspect, the embodiment of the present application provides a communication device, the communication device includes a processor and a memory, and a computer program is stored in the memory; the processor executes the computer program stored in the memory, so that the communication device executes The method described in the first aspect above.

In an eighth aspect, the embodiment of the present application provides a communication device, the communication device includes a processor and a memory, and a computer program is stored in the memory; the processor executes the computer program stored in the memory, so that the communication device executes The method described in the second aspect above.

In the ninth aspect, the embodiment of the present application provides a communication device, the device includes a processor and an interface circuit, the interface circuit is used to receive code instructions and transmit them to the processor, and the processor is used to run the code instructions to make the The device executes the method described in the first aspect above.

In the tenth aspect, the embodiment of the present application provides a communication device, the device includes a processor and an interface circuit, the interface circuit is used to receive code instructions and transmit them to the processor, and the processor is used to run the code instructions to make the The device executes the method described in the second aspect above.

In the eleventh aspect, the embodiment of the present application provides a channel state information CSI compression feedback system, the system includes the communication device described in the third aspect and the communication device described in the fourth aspect, or, the system includes the communication device described in the fifth aspect The communication device described in the above aspect and the communication device described in the sixth aspect, or, the system includes the communication device described in the seventh aspect and the communication device described in the eighth aspect, or, the system includes the communication device described in the ninth aspect And the communication device described in the tenth aspect.

In the twelfth aspect, the embodiment of the present invention provides a computer-readable storage medium, which is used to store the instructions used by the above-mentioned terminal equipment, and when the instructions are executed, the terminal equipment executes the above-mentioned first aspect. method.

In a thirteenth aspect, an embodiment of the present invention provides a readable storage medium for storing instructions used by the above-mentioned network equipment, and when the instructions are executed, the network equipment executes the method described in the above-mentioned second aspect .

In a fourteenth aspect, the present application further provides a computer program product including a computer program, which, when run on a computer, causes the computer to execute the method described in the first aspect above.

In a fifteenth aspect, the present application further provides a computer program product including a computer program, which, when run on a computer, causes the computer to execute the method described in the second aspect above.

In the sixteenth aspect, the present application provides a chip system, the chip system includes at least one processor and an interface, used to support the terminal device to realize the functions involved in the first aspect, for example, determine or process the data involved in the above method and at least one of information. In a possible design, the chip system further includes a memory, and the memory is used to store necessary computer programs and data of the terminal device. The system-on-a-chip may consist of chips, or may include chips and other discrete devices.

In the seventeenth aspect, the present application provides a chip system, the chip system includes at least one processor and an interface, used to support the network device to realize the functions involved in the second aspect, for example, determine or process the data involved in the above method and at least one of information. In a possible design, the chip system further includes a memory, and the memory is used for saving necessary computer programs and data of the network device. The system-on-a-chip may consist of chips, or may include chips and other discrete devices.

In an eighteenth aspect, the present application provides a computer program that, when run on a computer, causes the computer to execute the method described in the first aspect above.

In a nineteenth aspect, the present application provides a computer program that, when run on a computer, causes the computer to execute the method described in the second aspect above.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiment of the present application or the background art, the following will describe the drawings that need to be used in the embodiment of the present application or the background art.

FIG. 1 is a schematic structural diagram of a communication system provided by an embodiment of the present application;

FIG. 2 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application;

FIG. 3 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 4 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application;

FIG. 5 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 6 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 7 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 8 is a schematic flow chart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 9 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 10 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 11 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 12 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 13 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 14 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 15 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

FIG. 16 is a schematic flowchart of a channel state information CSI compression feedback method provided by an embodiment of the present application;

Fig. 17 is a schematic structural diagram of a communication device provided by an embodiment of the present application;

Fig. 18 is a schematic structural diagram of another communication device provided by an embodiment of the present application;

FIG. 19 is a schematic structural diagram of a chip provided by an embodiment of the present application.

Detailed ways

For ease of understanding, terms involved in this application are firstly introduced.

1. Channel State Information (CSI)

In the field of wireless communications, CSI is the propagation characteristic of a communication link. It describes the attenuation factor of the signal on each transmission path in the communication link, that is, the value of each element in the channel gain matrix, such as signal scattering, environmental attenuation, distance attenuation and other information. CSI can make the communication system adapt to the current channel conditions, and provides a guarantee for high-reliability and high-speed communication in a multi-antenna system.

The channel state information is divided into channel state information on the transmitter side and channel state information on the receiver side according to different application locations. Generally speaking, the channel state information on the transmitter side can compensate for fading in advance by means of power allocation, beamforming, and antenna selection to complete high-speed and reliable data transmission.

2. Massive Multiple-Input Multiple-Output (MIMO)

The mMIMO technology refers to the use of multiple transmitting antennas and receiving antennas at the transmitting end and the receiving end, respectively, so that signals are transmitted and received through multiple antennas at the transmitting end and the receiving end, thereby improving communication quality. mMIMO can make full use of space resources and double the system communication capacity without increasing spectrum resources and antenna transmission power.

mMIMO technology has been applied in 4G communication, and will be applied more deeply in 5G communication. MIMO in 4G communication has up to 8 antennas, and 16/32/64/128 or even larger antennas will be realized in 5G.

The mMIMO technology has the following advantages: High multiplexing gain and diversity gain: Compared with the existing MIMO system, the spatial resolution of the massive MIMO system is significantly improved. On the same time-frequency resource, use the spatial freedom provided by massive MIMO to communicate with the base station at the same time, and improve the multiplexing capability of spectrum resources between multiple terminal devices, thereby greatly improving the density and bandwidth of the base station without increasing the base station density and bandwidth. Spectral efficiency. High energy efficiency: The massive MIMO system can form narrower beams and radiate them in a smaller space area, so that the energy efficiency of the radio frequency transmission link between the base station and the terminal equipment is higher, and the transmission power loss of the base station is reduced , is an important technology for building future energy-efficient green broadband wireless communication systems. High spatial resolution: Massive MIMO systems have better robust performance. Since the number of antennas is far greater than the number of terminal devices, the system has a high degree of spatial freedom and a strong anti-interference capability. When the number of base station antennas tends to infinity, the negative effects such as additive white Gaussian noise and Rayleigh fading are all negligible

In order to better understand the method for compressing and feeding back channel state information CSI disclosed in the embodiment of the present application, the communication system to which the embodiment of the present application applies is firstly described below.

Please refer to FIG. 1 . FIG. 1 is a schematic structural diagram of a communication system provided by an embodiment of the present application. The communication system may include, but is not limited to, a network device and a terminal device. The number and form of the devices shown in Figure 1 are for example only and do not constitute a limitation to the embodiment of the application. In practical applications, two or more network equipment, two or more terminal equipment. The communication system shown in FIG. 1 includes one network device 101 and one terminal device 102 as an example.

It should be noted that the technical solutions of the embodiments of the present application may be applied to various communication systems. For example: long term evolution (LTE) system, fifth generation (5th generation, 5G) mobile communication system, 5G new radio (new radio, NR) system, or other future new mobile communication systems, etc. It should also be noted that the side link in this embodiment of the present application may also be referred to as a side link or a through link.

The network device 101 in the embodiment of the present application is an entity on the network side for transmitting or receiving signals. For example, the network device 101 may be an evolved base station (evolved NodeB, eNB), a transmission point (transmission reception point, TRP), a next generation base station (next generation NodeB, gNB) in the NR system, or a base station in other future mobile communication systems Or an access node in a wireless fidelity (wireless fidelity, WiFi) system, etc. The embodiment of the present application does not limit the specific technology and specific device form adopted by the network device. The network device provided by the embodiment of the present application may be composed of a centralized unit (central unit, CU) and a distributed unit (distributed unit, DU), wherein the CU may also be called a control unit (control unit), using CU-DU The structure of the network device, such as the protocol layer of the base station, can be separated, and the functions of some protocol layers are placed in the centralized control of the CU, and the remaining part or all of the functions of the protocol layer are distributed in the DU, and the CU centrally controls the DU.

The terminal device 102 in the embodiment of the present application is an entity on the user side for receiving or transmitting signals, such as a mobile phone. The terminal equipment may also be called terminal equipment (terminal), user equipment (user equipment, UE), mobile station (mobile station, MS), mobile terminal equipment (mobile terminal, MT) and so on. The terminal device can be a car with communication functions, a smart car, a mobile phone, a wearable device, a tablet computer (Pad), a computer with a wireless transceiver function, a virtual reality (VR) terminal device, an augmented reality ( augmented reality (AR) terminal equipment, wireless terminal equipment in industrial control (industrial control), wireless terminal equipment in self-driving (self-driving), wireless terminal equipment in remote medical surgery (remote medical surgery), smart grid ( Wireless terminal devices in smart grid, wireless terminal devices in transportation safety, wireless terminal devices in smart city, wireless terminal devices in smart home, etc. The embodiment of the present application does not limit the specific technology and specific device form adopted by the terminal device.

With the development of the fifth generation wireless communication network, mMIMO has become a key technology. By configuring a large number of antennas, mMIMO not only greatly improves the channel capacity under limited spectrum resources, but also has a strong anti-interference capability. In order to make better use of the mMIMO technology, the transmitter needs to obtain CSI. In the system, the UE side estimates the CSI of the downlink channel, and then feeds the CSI back to the BS through a feedback link with a fixed bandwidth. However, due to the multi-antenna nature of mMIMO, the overhead of CSI feedback is huge, so how to efficiently and accurately feed back CSI is still a serious challenge.

In order to reduce the feedback overhead of CSI, researchers have proposed many algorithms based on compression estimation theory, most of which use the spatial and temporal correlation of the channel to reduce the feedback overhead. In the traditional compression method, the CS-based feedback method transforms the CSI matrix into a sparse matrix under a certain basis, and uses the method in the computer field for feedback. The quantization-based codebook compression method quantizes the CSI into a certain number of bits. In recent years, with the rapid development of deep learning, deep learning has been widely used in computer vision, speech signal processing and natural language processing and other fields. Due to the powerful parallel computing, adaptive learning and cross-domain knowledge sharing capabilities of deep learning networks, deep learning methods are gradually being applied in the field of CSI compression feedback to further reduce CSI feedback overhead. For example, a deep learning network regards MIMO channel data as image information, then uses an encoder to compress CSI, and finally uses a decoder to restore it to achieve the purpose of CSI feedback. An improved deep learning network for CSI compression and feedback by exploiting the temporal correlation of channels.

In related technologies, the CSI feedback method based on channel spatial correlation uses a correlation algorithm to divide channel elements with spatial correlation into several clusters, and maps multiple channel elements in each cluster into a single representation value, At the same time, it will be divided into several group modes according to the different cluster division methods. The selected group mode and characterization value are fed back to the transmitter through a feedback link for CSI reconstruction. However, this method requires strong spatial correlation between channel elements, and cannot achieve accurate CSI compression and feedback for channels with little spatial correlation. Moreover, the algorithm complexity of this method is high, and as the number of antennas at the transmitting end increases, the number of clusters increases, and the feedback overhead is still huge.

In the related art, the CSI matrix in the space-frequency domain can also be transformed into the CSI matrix in the angle domain through a two-dimensional discrete Fourier transform (Discrete Fourier Transform, DFT). The real and imaginary parts of the CSI matrix are then separated to obtain two-dimensional CSI image information. In the angle domain, due to the delay of multipath arrival and the sparsity of mMIMO channel information matrix, the main value part of the CSI image is extracted. Then the extracted CSI matrix is used as the input of the deep learning network for training, where the encoder is deployed on the UE side to compress the extracted CSI image into a low-dimensional codeword, and the decoder is deployed on the BS side to convert The compressed low-dimensional codeword is restored to the corresponding CSI image, and the reconstructed channel is obtained. Finally, offline training and parameter updating are performed on the deep learning network, so that the reconstructed channel is as close as possible to the channel in the original angle domain. Finally, the inverse two-dimensional DFT transform is performed on the reconstructed channel to obtain the CSI matrix in the original space-frequency domain. Apply the trained deep learning network model to online deployment applications.

However, the above deep learning-based CSI compression and feedback method only uses the original channel parameters for compression and feedback. For time-series CSI with time correlation, the original channel parameters cannot well reflect the structural characteristics and time-correlation characteristics of time-series CSI. . Moreover, the above method only compresses the CSI from the perspective of the image, and it is impossible to accurately compress and reconstruct the CSI by using the time-correlation feature for the time-series CSI with time correlation.

It can be understood that the communication system described in the embodiment of the present application is to illustrate the technical solution of the embodiment of the present application more clearly, and does not constitute a limitation to the technical solution provided in the embodiment of the present application. With the evolution of the system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

The method and device for compressing and feeding back channel state information (CSI) provided by the present application will be described in detail below with reference to the accompanying drawings.

Please refer to FIG. 2 . FIG. 2 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 2, the method is applied to a terminal device, and the method may include but not limited to the following steps:

Step 201: Obtain an estimated CSI image H of a network device, and generate a time-series CSI image _Hc according to the estimated CSI image H.

In the embodiment of the present application, in the communication between the terminal device and the network device, the frequency division duplex (Frequency Division Duplex, FDD) method is used to transmit signals, and the frequency division multiplexing technology is used to separate the transmitted and received signals. The upload and download segments are separated by a "frequency offset". In the mMIMO downlink of FDD, that is, in the channel from the network device to the terminal device, multiple antennas are deployed on the network device, and Orthogonal Frequency Division Multiplexing (OFDM) technology is used for mMIMO transmission. , with multiple subcarriers in the channel. In order to transmit signals better and improve the performance of the mMIMO system, the network device, as the sender of the downlink, needs to obtain CSI. The CSI is obtained by channel estimation by the terminal device, and the size of the estimated CSI image H is T×c×N _s ×N _t . In this application, the CSI acquired by the terminal device has time correlation, that is, the CSI is time sequence. Wherein, T is the length of the time sequence; c is the dimension of the real and imaginary parts of the channel, and the channel has only one real part and one imaginary part, that is, c=2; N _s is the number of subcarriers; N _t is the network device The number of deployed antennas.

The estimated CSI image H contains spatial domain information. In order to transform the estimated CSI image H into an angular domain, two-dimensional DFT is performed on the estimated CSI image H, and the time-series CSI image _Hc can be obtained after transformation. The _Hc is more sparse than H. Due to the influence of multipath time delay, the transformed estimated CSI image H only has values in the first _Nc rows, and the _Nc is the number of effective rows, so only the first _Nc rows are reserved. data, so the size of the H _c is T×c×N _c ×N _t .

In a possible embodiment, N _t =32 antennas are configured in a linear antenna array (Uniform Linear Array, ULA) manner at half-wavelength intervals on the network device, and a single antenna is configured on the terminal device. Using the COST2100 channel model, 150,000 space-frequency domain CSI matrix samples were generated in the 5.3GHz indoor micro-cell scene, and divided into a training set with 100,000 samples, a validation set with 30,000 samples, and a test set with 20,000 samples. The mMIMO system adopts OFDM technology, and the subcarrier N _s =1024, then the CSI information matrix in the space-frequency domain is

And the time sequence length T of each CSI information matrix is set as T=5. In this case, the number of feedback parameters of each CSI information matrix is 1024×32, and the overhead is huge under the premise of limited feedback bandwidth.

In order to reduce the feedback overhead, the CSI information matrix H is transformed from the space-frequency domain to the angle-delay domain by using two-dimensional DFT, namely

F _d and

are discrete Fourier transform matrices with sizes of 1024×1024 and 32×32 respectively, and the superscript H indicates the conjugate transpose of the matrix. In the angle delay domain, using the delay of multipath arrival in a finite time period, we intercept the main value part of the first N _c =32 rows of H _c . At this time, the size of the time-series CSI image H _c in the angle domain is 5 ×2×32×32.

Step 202: Compress the time-series CSI image _Hc to generate a feature codeword;

In the embodiment of the present application, it is necessary to feed back the time-series CSI image H _c to the network device, and sending the H _c directly causes too large channel resources and wastes resources. The time-series CSI image H _c needs to be compressed and simplified to save resources. In the embodiment of the present application, the time-series CSI image is projected to the self-information domain through the self-information domain converter, so as to obtain the time-series self-information image _He . Due to the difference in information carried by each part of the high-frequency channel image, projecting the time-series CSI image into the dimension of self-information can highlight its structural characteristics and temporal correlation characteristics, and the time-series CSI image has better reliability under the dimension of self-information. Compressibility.

Then, input the time-series self-information image He _into the time-series feature coupling encoder, use the cyclic neural network LSTM to extract the time correlation information between the self-information images, and use the one-dimensional space compression network to obtain the channel image projected on the self-information domain Structural feature information, adding and coupling the extracted time correlation information and structural feature information to obtain the final implicit feedback feature codeword.

Step 203: Send the feature code word to the network device.

In the embodiment of the present application, the feature codeword is obtained after compressing the time-series CSI image _Hc . The feature codeword includes relevant information of the time-series CSI image _Hc . After the feature codeword is sent to the network device, the network device restores the feature codeword to obtain a restored time-series CSI image, and performs mMIMO transmission according to the restored CSI image.

By implementing the embodiment of the present application, the terminal device can compress the time-series CSI image H _c corresponding to the estimated CSI image H to generate a feature codeword, and feed back the time-series CSI image to the network device through the feature codeword. The channel resource occupied by the feedback CSI image can be reduced, resources are saved, and the accuracy of the feedback CSI image is improved.

Please refer to FIG. 3 . FIG. 3 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 3, the method may include but not limited to the following steps:

Step 301: Input the time-series CSI image H _c into the self-information domain converter to generate a time-series self-information image _He , wherein the time-series CSI image H _c and the time-series self-information image He _e both have a time dimension of T ;

In the embodiment of the present application, the time-series CSI image H _c is projected into the self-information domain through the self-information domain converter to obtain the time-series self-information image _He . Due to the difference in information carried by each part of the high-frequency channel image, the Projecting time-series CSI images to the dimension of self-information can highlight its structural features and temporal correlation features, and time-series CSI images have better compressibility in the dimension of self-information.

Step 302: Input the time-series self-information image _He into a time-series feature coupling encoder for feature extraction to generate a structural feature matrix and a temporal correlation matrix;

In the embodiment of the present application, the structural features and time correlation features of the time series self-information image _He are extracted through a time series feature coupling encoder, and the structural feature matrix includes the structural features and the time correlation matrix includes all The time-dependent characteristics described above.

Step 303: Generate the feature code word according to the structure feature matrix and the time correlation matrix.

In the embodiment of the present application, the structural feature matrix and the temporal correlation matrix are coupled to obtain the feature codeword.

Please refer to FIG. 4 . FIG. 4 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 4, the method may include but not limited to the following steps:

Step 401a: Input the time-series CSI image _Hc into the three-dimensional convolutional feature extraction network to extract features to obtain the first time-series feature image F, wherein the convolution kernel specification of the three-dimensional convolutional network is f×t×n×n , the f is the number of features extracted, the t is the depth of the convolution in the time dimension, and the n is the length and width of the convolution window;

In the embodiment of the present application, the time-series CSI image H _c contains information of the time dimension, so the two-dimensional convolutional layer cannot effectively extract the features in it. The embodiment of the present disclosure uses a three-dimensional feature convolution feature extraction network to extract Features in the time-series CSI image. The three-dimensional convolutional network includes a convolutional layer, a three-dimensional normalization layer, and an activation function layer. The convolution kernel specification of the convolutional layer in the three-dimensional convolutional network is f×t×n×n, that is, the convolution kernel is from Each convolution in the time series CSI image H _c will extract f features; in order to prevent gradient disappearance or gradient explosion, the output of the convolution layer is input into the three-dimensional normalization layer for normalization, and finally the activation function layer to obtain the first time-series feature image F, the

The activation function of the activation function layer is a LeakyReLU activation function, and the LeakyReLU activation function is formulated as follows:

In a possible embodiment, the three-dimensional convolutional feature extraction network converts the time-series CSI image _Hc into the first time-series feature image F∈R ^{5×64×32×32} , where each CSI image extracts 64 features, Corresponds to dimension 64. Since the time-series CSI image contains the time dimension, the two-dimensional convolutional layer cannot effectively extract its features. Therefore, the feature extraction network in the present invention uses a three-dimensional convolutional layer, and the size of the convolution kernel is 64×1×3×3.

Step 401b: Generate a first index matrix M according to the time-series CSI image _Hc ;

In the embodiment of the present application, when the time-series CSI image _Hc is input into the three-dimensional convolutional feature extraction network, it needs to be input into the self-information module to extract self-information, which can be used to measure the information contained in a single event. The amount is large, and the self-information image is obtained according to the self-information; and the second index matrix is obtained by mapping the self-information image through the index matrix module.

Step 402: Obtain a time-series self-information image _He according to the first time-series feature image F and the first index matrix M.

In the embodiment of the present application, after the first time-series feature image F and the first index matrix M are obtained, the first time-series feature image F and the first index matrix M are multiplied point-to-point to obtain the removal information The redundant information feature image, that is, the second information feature image, and perform dimension reduction on the second information feature image to generate the time-sequence self-information image _He .

Please refer to FIG. 5 . FIG. 5 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 5, the method may include but not limited to the following steps:

Step 501: Input the time-series CSI image _Hc into the self-information module to generate self-information of the area to be estimated in the time-series CSI image _Hc , and use it as a self-information image;

In the embodiment of the present application, the time-series CSI image _Hc contains information of the time dimension, that is, time series, and it is necessary to calculate the self-information in the time-series CSI image Hc at each time point in the time series, and calculate the time-series CSI image _Hc at each time point The self-information of is composed of a corresponding self-information image.

Step 502: Input the self-information image into an index matrix module for mapping to obtain a first index matrix M.

In the embodiment of the present application, after the self-information image is acquired, the self-information image is input into the index matrix module. The index matrix module includes a mapping network, a decision device and a splicing module. The self-information image is mapped to the self-information domain by a mapping module to obtain a second index matrix. The second index matrix corresponds to the time-series CSI image H _c at each time point in the time series. In order to maintain the information of the time dimension, the second index matrix needs to be spliced in the order of the time series to obtain the first index matrix M .

Please refer to FIG. 6 . FIG. 6 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 6, the method may include but not limited to the following steps:

Step 601: Split the time-series CSI image H _c in time series to obtain split images H _c,i at each time point;

In the embodiment of the present application, the time-series CSI image H _c includes information of a time dimension, that is, a time series, and self-information in the time-series CSI image H _c at each time point in the time series needs to be calculated. Split the time-series CSI image H _c according to the time sequence, and obtain the split image H _c,i at each time point,

i∈(1,2,···,T), there are T dimensions in the time series, so splitting will obtain T split images H _c,i .

Step 602: Divide the split image into multiple regions p _j to be estimated, and obtain self-information estimation values of the regions to be estimated

According to the estimated value from the self-information

Generated from the information image I _c,i .

In the embodiment of the present application, after obtaining the split image H _c,i , for each H _c,i , the real part is used

Indicates that the imaginary part is

express. We use n×n panes to

and

Divided into multiple regions to be estimated, each region to be estimated is represented by p _j ∈ ^{R n×n} , j∈[1,2,···,(N _c -n+1)(N _t -n+ 1)]. The self-information calculation formula of each area p _j is as follows:

in,

is the estimated value of self-information of p _j ; N _j is the set of all areas near p _j ; p′ _j,r is the area near the rth of p _j ; r=[1,2,···,(2R +1) ² ]; R is the radius of Manhattan, used to determine the boundary of N _j ; h is the bandwidth, used to adjust the influence of the distance between p _j and p′ _{j, r} on the calculation of self-information, and constant is a constant.

In a possible embodiment, the split CSI image is expressed as H _c,i ∈R ^2×32×32 , i=(1,2,···,5). For each H _c,i , the real part uses

Indicates that the imaginary part is

express. We use a 1×1 pane to

and

Divided into multiple regions, each region is represented by p _j ∈ ^{R 1×1} , j=[1,2,···,1024]. where in p′ _j,r , r=[1,2,···,49]. Manhattan radius R=3, bandwidth h=1, and constant=3×10 ^-6 . In order to simplify the calculation, we use each pixel in H _c,i as a region to calculate the self-information, and finally all the self-information values

Compose the matrix to get the self-information matrix I _c,i ∈R ^2×32×32 of H _c,i .

In a possible embodiment, in order to simplify the calculation, each pixel in H _c,i is regarded as a region to calculate the self-information estimated value of p _j

All of the self-information values

Form a matrix to get the self-information image of H _c,i

Please refer to FIG. 7 . FIG. 7 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 7, the method may include but not limited to the following steps:

Step 701: Input the self-information image into the mapping network to extract features to obtain a first information feature image D _c,i , wherein the mapping network is a two-dimensional convolutional neural network;

In the embodiment of the present application, the mapping network includes a two-dimensional convolutional layer, a two-dimensional normalization layer and an activation function layer. The self-information image contains information that only contains two dimensions, so the size of the convolution kernel in the two-dimensional convolutional layer is f×n×n. After extracting features through the two-dimensional convolutional layer, input the two-dimensional The normalization layer normalizes the feature values, and finally inputs the activation function layer to obtain the first information feature image D _c,i , and the activation function of the activation function layer is a LeakyReLU activation function.

Step 702: Input the first information characteristic image D _c,i into the decision device for binarization processing to obtain a second index matrix M _i ;

In the embodiment of the present application, the first information feature image D _c,i is binarized by the decision unit, and the threshold Y is set by the decision unit, and the first information feature image D _c,i The element value in each element of the element, if the element value is greater than or equal to the threshold Y set by the decision maker, the corresponding element of the element value is set to 1; if the element value is smaller than the threshold Y set by the decision maker , then set the corresponding element of the element value to 0. to obtain the second index matrix M _i , where,

In a possible embodiment, the threshold Y=9.288, the decider sets the positions of elements less than 9.288 in D _{c,i to} 0, and sets the positions of elements greater than 9.288 to 1 to obtain the final index matrix M _i ∈ R ^64×32×32

Step 703: Concatenate the second index matrix M _i to obtain a first index matrix M.

In the embodiment of the present application, the split image H _c,i corresponding to the second index matrix M _i is an image at a time point in the time-series CSI image H _c . Therefore, the second index matrix M _i may be concatenated according to the order of the time series to obtain the first index matrix M.

Please refer to FIG. 8 . FIG. 8 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 8, the mapping network includes a two-dimensional convolutional layer, a two-dimensional normalization layer and an activation layer, and the method may include but not limited to the following steps:

Step 801: Input the self-information image into the two-dimensional convolution layer to extract features, so as to obtain a first feature image;

In the embodiment of the present application, the self-information image contains information of only two dimensions, so the size of the convolution kernel in the two-dimensional convolution layer is f×n×n, and the two-dimensional convolution layer extracts features to get the first feature image.

Step 802: Input the first feature image into the two-dimensional normalization layer to normalize the pixel values in the first feature image to obtain a second feature image;

In the embodiment of the present application, in order to prevent gradient disappearance and gradient explosion, the first feature image is input into the two-dimensional normalization layer, and the value of each pixel in the second feature image is normalized, Make the magnitude of the pixel value in the range [0, 1].

Step 803: Input the second feature image into an activation function layer for nonlinear mapping to obtain the first information feature image D _c,i .

In the embodiment of the present application, the activation function of the activation function layer adopts the LeakyReLU activation function.

In a possible embodiment, the mapping network maps I _c,i to the information feature image D _c,i ∈ R ^64×32×32 , and the size of the convolution kernel of the two-dimensional convolutional layer in the mapping network is 64×3 ×3.

Please refer to FIG. 9 . FIG. 9 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 9, the method may include but not limited to the following steps:

Step 901: multiply the first time-series feature image F by the first index matrix M to obtain a second information feature image;

In the embodiment of the present application, the first time-series feature image F and the first index matrix M are obtained by removing information redundancy through the three-dimensional convolutional feature extraction network, the self-information module, and the index matrix module, Multiplying the first time-series feature image F and the first index matrix M to obtain a second information feature image, where information features in the second information feature image are more refined and can better reflect channel features.

Step 902: Input the second information feature image into a dimension restoration network to perform dimension restoration, so as to generate the time-series self-information image _He .

In the embodiment of the present application, the dimension reduction network includes a three-dimensional convolutional layer, a three-dimensional normalization layer and an activation function layer. The size of the convolution kernel in the three-dimensional convolutional layer is c×t×n×n, wherein c is the restored dimension, and c is the dimension of virtual and real parts, so c=2; t is the depth of convolution in the time dimension; n×n is the specification of the convolution window, that is, the length and width of the convolution window are both n. The three-dimensional normalization layer performs normalization processing on the output of the three-dimensional convolutional layer, and the activation function of the activation function layer is a LeakyReLU activation function. Dimensional reduction is performed on the second information characteristic image through the dimension reduction network to obtain the time-series self-information image _He . Compared with the time-series CSI image _Hc, the time-series self-information image _He has more obvious structural features and temporal correlation features.

Optionally, the size of the convolution kernel of the three-dimensional convolution layer in the dimension reduction network is 2×1×3×3.

Optionally, the temporal feature coupled encoder includes a one-dimensional space-time compression network and a coupled long short-term memory network (Long Short Term Memory Network, LSTM).

The time sequence is input into the one-dimensional space-time compression network for one-dimensional space-time compression after dimension transformation from the information image _He to obtain the structural feature matrix, wherein the convolution kernel specification of the one-dimensional space-time compression network is S× 2N _c N _t ×m, the 2N _c N _t is the length of the convolution window, the m is the width of the convolution window, S is the target dimension, and the dimension of the structural feature matrix is T×S. Among them, S=cN _c N _t /σ, σ is the compression ratio.

In a possible embodiment, the size of the convolution kernel of the one-dimensional convolutional layer in the one-dimensional space-time compression network is S×1. The one-dimensional space compression network compresses the time-series self-information image _He into S-dimensional structural feature information according to the compression rate, where S=2048/σ.

In the embodiment of the present application, in order to save channel resources, the terminal device feeds back the timing self-information image _He to the network device in the form of a codeword. In order to convert the time-series self-information image _He into a feature codeword, extract the structural features and temporal correlation features of the time-series self-information image _He through the time-series feature coupling encoder, generate the structural feature matrix and the The time correlation matrix described above.

Please refer to FIG. 10 . FIG. 10 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 10, the method may include but not limited to the following steps:

Step 1001: Transform the time-series self-information image H _e into a coupled LSTM to extract features, so as to obtain the time-correlation matrix, wherein the dimension of the time-correlation matrix is T×S;

In the embodiment of the present application, the time-series self-information image _He is dimensionally transformed and then input into LSTM. The LSTM contains multiple structural units and is suitable for processing and predicting important events with very long intervals and delays in time series. The temporal correlation feature is extracted by the coupled LSTM to generate the temporal correlation matrix. The dimension of the temporal correlation matrix is the same as that of the structural feature matrix, both being T×S.

In a possible embodiment, the number of structural units in the LSTM is T, which is equal to the time dimension of the time-series self-information image He _e . The structural units are connected in series, and the output of one structural unit is input to the next structural unit.

Step 1002: Coupling the structural feature matrix and the temporal correlation feature matrix to generate the feature codeword.

In the embodiment of the present application, the dimensions of the structural feature matrix and the time-correlation feature matrix are the same, and the values of the corresponding points in the structural feature matrix and the time-correlation feature matrix are added for coupling to generate the The above feature codeword.

Please refer to FIG. 11 . FIG. 11 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 11, the method may include but not limited to the following steps:

Step 1101: Input the training time-series CSI image _Hc into the self-information domain converter to obtain the training time-series self-information image _He ;

In the embodiment of the present application, it is necessary to train the self-information domain converter and the temporal feature coupling encoder in the terminal device, and input the training time-series CSI image H _c from the information domain converter to obtain the training time-series self- Information image _He .

Step 1102: Input the training time-series self-information image He _e into a time-series feature coupling encoder to obtain training feature codewords.

In the embodiment of the present application, the training time-series self-information image He _is input into the time-series feature coupling encoder to obtain the training feature codeword, and conduct preliminary training to obtain the self-information domain converter and the time-series feature coupling code network parameters in the device.

Optionally, also include:

In the embodiment of the present application, in order for the network device to successfully decode the information in the feature codeword, the training data needs to be sent to the network device to adjust the network parameters of the decoupling module in the network device .

Please refer to FIG. 12 . FIG. 12 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 12, the method is applied to a network device, and the method may include but not limited to the following steps:

Step 1201: Receive the feature code word sent by the terminal device;

In the embodiment of the present application, the network device is used as a downlink sending end, and in order to obtain better signal transmission and improve the performance of the mMIMO system, the network device needs to obtain CSI. The time-series CSI image is restored according to the feature code word sent by the terminal device.

Step 1202: Restoring the feature codewords to obtain restored time-series CSI images

In the embodiment of the present application, the terminal device restores the feature codeword through the decoupling module and the restored convolutional neural network to obtain the restored time-series CSI image

The decoupling module includes a one-dimensional space-time decompression network and a decoupling LSTM. The restored time-series CSI image

The size and the size of the time-series CSI image H _c are both T×c×N _c ×N _t .

Step 1203: Restoring the time series CSI image according to the

Get the restored estimated CSI image

In the embodiment of this application, for the restored time series CSI image

Perform two-dimensional DFT inverse transformation to obtain the restored and estimated CSI image required by network equipment

By implementing the embodiment of the present application, the network device decompresses the characteristic code word compressed by the terminal device to obtain the restored restored estimated CSI image

Optionally, the restoring the feature codeword includes:

Inputting the feature codeword into a time-series feature decoupling decoder to obtain the restored time-series CSI image

Please refer to FIG. 13 . FIG. 13 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in FIG. 13, the sequential feature decoupling decoder includes a decoupling module and a restored convolutional neural network. The method may include but not limited to the following steps:

Step 1301: Input the feature code word into the decoupling module for decoupling, so as to obtain the restored time series self-information image

In the embodiment of the present application, the decoupling module includes a one-dimensional space-time decompression network and a decoupling LSTM. To the structural feature information in the feature codeword, the decoupling LSTM is used to extract the time correlation information in the feature codeword.

Step 1302: Restore the time series from the information image

In the embodiment of the present application, the restored convolutional neural network is used to restore time-series self-information images

Recover the corresponding restored timing CSI image

Please refer to FIG. 14 . FIG. 14 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 14, the decoupling module includes a one-dimensional space-time decompression network and a decoupling LSTM, and the method may include but not limited to the following steps:

Step 1401: Input the feature codeword into the one-dimensional space-time decompression network for decompression to obtain the restored structure feature matrix;

In the embodiment of the present application, the one-dimensional space-time decompression network includes a one-dimensional convolutional layer, and the one-dimensional convolutional layer includes a plurality of one-dimensional convolutional kernels.

Step 1402: Input the feature codeword into the decoupling LSTM for decoupling, so as to obtain the restored time correlation matrix;

Step 1403: Obtain the restored time-series self-information image according to the restored structural feature matrix and the restored temporal correlation matrix

In the embodiment of the present application, the dimensions of the restored structural feature matrix and the restored time correlation matrix are the same, both being T×cN _c N _t . The restored time-series self-information image can be obtained by adding the restored structure feature matrix and the restored time correlation matrix point-to-point and performing dimension transformation

Optionally, the convolution kernel specification of the one-dimensional space-time decompression network is 2N _c N _t × S × m, the T is the number of rows of the restored temporal correlation matrix, and the 2N _c N _t is the Describes the number of columns of the restored time dependence matrix.

Optionally, the acquiring the restored time-series self-information image according to the restored structural feature matrix and the restored time correlation matrix includes:

In a possible embodiment, the convolution kernel of the first convolution layer is 8×1×3×3, the convolution kernel of the second convolution layer is 16×1×3×3, and the convolution kernel of the third convolution layer is The product kernel is 2×1×3×3, the convolution kernel of the fourth convolution layer is 8×1×3×3, the convolution kernel of the fifth convolution layer is 16×1×3×3, and the sixth convolution layer The convolution kernel is 2×1×3×3, the step size of the first six three-dimensional convolution layers is 1, and the activation function uses the LeakyReLU function. The seventh convolutional layer is a normalization layer with a convolution kernel of 2×1×3×3 and a step size of 1.

Optionally, the restoring the time series from the information image

include:

Restore the time series from the information image

sum to obtain a fourth reduced feature map;

In the embodiment of the present application, in order to prevent the disappearance of the gradient during the training of the reconstructed CSI module, a short-circuit operation is performed on the first and fourth layers, and the fourth and sixth three-dimensional convolutional layers. The first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer have a step size of k, and each of the convolutional layers The activation function adopts the Sigmoid function, and the formula of the Sigmoid function is expressed as

Please refer to FIG. 15 . FIG. 15 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 15, the method may include but not limited to the following steps:

Step 1501: Receive the training data sent by the terminal device, the training data includes the training feature codeword, the time sequence length of the time-series self-information image _He , the dimension of the training feature codeword and the training time-series CSI image;

In the embodiment of the present application, it is necessary to train the decoupling module and the restored convolutional neural network in the time series feature decoupling decoder deployed by the terminal device. Training is performed according to the training data sent by the terminal device.

Step 1502: Obtain a restored time-series CSI image according to the training feature codeword;

In the embodiment of the present application, the training feature codeword is input into the time-series feature decoupling decoder for restoration, so as to obtain the restored CSI image.

Step 1503: Perform training according to the restored time-series CSI images and the training time-series CSI images.

In the embodiment of the present application, it is necessary to optimize the network parameters in the timing feature decoupling decoder so that the restored timing CSI image and the training timing CSI image are as close as possible.

Please refer to FIG. 16 . FIG. 16 is a schematic flowchart of a method for compressing and feeding back channel state information (CSI) provided by an embodiment of the present application. The method can be applied to various communication systems. For example: 5th generation (5th generation, 5G) mobile communication system, 5G new air interface (new radio, NR) system, or other future new mobile communication systems, etc. As shown in Figure 16, the method may include but not limited to the following steps:

Step 1601: Determine the number of structural units in the decoupled LSTM according to the time series length of the time series self-information image _He ;

In the embodiment of the present application, in order to improve the effect of decoupling, it is necessary to make the structure of the decoupled LSTM symmetrical to the structure of the coupled LSTM, and the number of structural units in the coupled LSTM is T, which is the same as the timing self-information The time series of images He and _e have the same length, so the number of structural units in the decoupled LSTM needs to be equal to T. The structural units in the decoupled LSTM are connected in series.

Step 1602: Determine the network parameters of the one-dimensional space-time decompression network according to the dimensions of the training feature codewords.

In the embodiment of the present application, the one-dimensional space-time decompression network has the same structure as the one-dimensional space-time compression network, and the feature number extracted by the one-dimensional space-time compression network is S, then the feature number decompressed by the one-dimensional space-time decompression network It should also be S, the dimension of the training feature codeword is T×S, then the size of the one-dimensional convolution kernel in the one-dimensional space-time decompression network is 2N _c N _t ×S×m.

Optionally, also include:

In the embodiment of this application, in the training process, in order to enable the network to learn the global optimal solution, the learning rate of the network adopts a "gradual learning" change method, and the learning rate increases linearly in the first few training cycles, and after reaching the peak value , the learning rate decreases slowly in a cosine trend, and the downward trend is as the formula expression of the above learning rate.

In a possible embodiment, the learning rate increases linearly in the first 30 learning cycles, and after reaching the peak, the learning rate decreases slowly in a cosine trend, where γ _max =2×10 ^-3 , γ _min =5×10 ^-5 , _Tw =30, T'=2000.

Optionally, also include:

In the embodiment of the present application, after the training is completed, the recommended network parameters of the decoupling module and the restored convolutional neural network can be obtained, and the decoupling module and the restored convolutional neural network are updated according to the recommended network parameters.

In a possible embodiment, the self-information domain converter of the terminal device is used to generate a time-series self-information image _He , and the time-series self-information image _{He is} input into a time-series feature coupling encoder to generate the feature codeword, Restore the feature codewords to a restored time-series CSI image by using a time-series feature coupling decoder in the network device

and then pass to

Perform two-dimensional DFT inverse transformation to obtain the restored estimated CSI image

During the process of transmitting the feature codeword, the network parameters are constantly updated.

Optionally, the method further includes: after the terminal device obtains the characteristic codeword through a time-series characteristic coupling encoder, for the convenience of transmission, the codeword is quantized by e bits and then fed back to the network device. Then use the trained network parameters to obtain the restored time series CSI image after the network device performs dequantization and time series feature decoupling decoder

Optionally, the codeword is quantized to 64 bits and then fed back to the network device.

In the foregoing embodiments provided in the present application, the methods provided in the embodiments of the present application are introduced from the perspectives of the network device and the first terminal device respectively. In order to realize the various functions in the method provided by the above-mentioned embodiments of the present application, the network device and the first terminal device may include a hardware structure and a software module, and realize the above-mentioned functions in the form of a hardware structure, a software module, or a hardware structure plus a software module . A certain function among the above-mentioned functions may be implemented in the form of a hardware structure, a software module, or a hardware structure plus a software module.

Please refer to FIG. 17 , which is a schematic structural diagram of a communication device 170 provided by an embodiment of the present application. The communication device 170 shown in FIG. 17 may include a transceiver module 1701 and a processing module 1702 . The transceiver module 1701 may include a sending module and/or a receiving module, the sending module is used to realize the sending function, the receiving module is used to realize the receiving function, and the sending and receiving module 1701 can realize the sending function and/or the receiving function.

The communication device 170 may be a terminal device (such as the first terminal device in the foregoing method embodiments), or a device in the terminal device, or a device that can be matched with the terminal device. Alternatively, the communication device 170 may be a network device, or a device in the network device, or a device that can be matched with the network device.

The communication device 170 is a terminal device:

An estimation module, configured to acquire an estimated CSI image H of the terminal device, and generate a time-series CSI image _Hc according to the estimated CSI image H;

A sending module, configured to send the feature codeword to a network device.

The communication device 170 is a network device:

Get the restored estimated CSI image

Please refer to FIG. 18 , which is a schematic structural diagram of another communication device 180 provided by an embodiment of the present application. The communication device 180 may be a network device, or a terminal device (such as the first terminal device in the foregoing method embodiments), or a chip, a chip system, or a processor that supports the network device to implement the above method, or a A chip, chip system, or processor that supports the terminal device to implement the above method. The device can be used to implement the methods described in the above method embodiments, and for details, refer to the descriptions in the above method embodiments.

Communications device 180 may include one or more processors 1801 . The processor 1801 may be a general purpose processor or a special purpose processor or the like. For example, it can be a baseband processor or a central processing unit. The baseband processor can be used to process communication protocols and communication data, and the central processing unit can be used to control communication devices (such as base stations, baseband chips, terminal equipment, terminal equipment chips, DU or CU, etc.) and execute computer programs , to process data for computer programs.

Optionally, the communication device 180 may further include one or more memories 1802, on which a computer program 1804 may be stored, and the processor 1801 executes the computer program 1804, so that the communication device 180 executes the method described in the foregoing method embodiments. method. Optionally, data may also be stored in the memory 1802 . The communication device 180 and the memory 1802 can be set separately or integrated together.

Optionally, the communication device 180 may further include a transceiver 1805 and an antenna 1806 . The transceiver 1805 may be called a transceiver unit, a transceiver, or a transceiver circuit, etc., and is used to implement a transceiver function. The transceiver 1805 may include a receiver and a transmitter, and the receiver may be called a receiver or a receiving circuit for realizing a receiving function; the transmitter may be called a transmitter or a sending circuit for realizing a sending function.

Optionally, the communication device 180 may further include one or more interface circuits 1807 . The interface circuit 1807 is used to receive code instructions and transmit them to the processor 1801 . The processor 1801 executes the code instructions to enable the communication device 180 to execute the methods described in the foregoing method embodiments.

In an implementation manner, the processor 1801 may include a transceiver for implementing receiving and sending functions. For example, the transceiver may be a transceiver circuit, or an interface, or an interface circuit. The transceiver circuits, interfaces or interface circuits for realizing the functions of receiving and sending can be separated or integrated together. The above-mentioned transceiver circuit, interface or interface circuit may be used for reading and writing code/data, or the above-mentioned transceiver circuit, interface or interface circuit may be used for signal transmission or transmission.

In an implementation manner, the processor 1801 may store a computer program 1803, and the computer program 1803 runs on the processor 1801, and may cause the communication device 180 to execute the methods described in the foregoing method embodiments. The computer program 1803 may be solidified in the processor 1801, and in this case, the processor 1801 may be implemented by hardware.

In an implementation manner, the communication device 180 may include a circuit, and the circuit may implement the function of sending or receiving or communicating in the foregoing method embodiments. The processors and transceivers described in this application can be implemented in integrated circuits (integrated circuits, ICs), analog ICs, radio frequency integrated circuits (RFICs), mixed-signal ICs, application specific integrated circuits (ASICs), printed circuit boards ( printed circuit board, PCB), electronic equipment, etc. The processor and transceiver can also be fabricated using various IC process technologies such as complementary metal oxide semiconductor (CMOS), nMetal-oxide-semiconductor (NMOS), P-type Metal oxide semiconductor (positive channel metal oxide semiconductor, PMOS), bipolar junction transistor (bipolar junction transistor, BJT), bipolar CMOS (BiCMOS), silicon germanium (SiGe), gallium arsenide (GaAs), etc.

The communication device described in the above embodiments may be a network device or a terminal device (such as the first terminal device in the foregoing method embodiments), but the scope of the communication device described in this application is not limited thereto, and the structure of the communication device can be Not limited by Figure 18. The communication means may be a stand-alone device or may be part of a larger device. For example the communication device may be:

(1) Stand-alone integrated circuits ICs, or chips, or chip systems or subsystems;

(2) A set of one or more ICs, optionally, the set of ICs may also include storage components for storing data and computer programs;

(3) ASIC, such as modem (Modem);

(4) Modules that can be embedded in other devices;

(5) Receivers, terminal equipment, intelligent terminal equipment, cellular phones, wireless equipment, handsets, mobile units, vehicle equipment, network equipment, cloud equipment, artificial intelligence equipment, etc.;

(6) Others and so on.

For the case where the communication device may be a chip or a chip system, refer to the schematic structural diagram of the chip shown in FIG. 19 . The chip shown in FIG. 19 includes a processor 1901 and an interface 1902 . Wherein, the number of processors 1901 may be one or more, and the number of interfaces 1902 may be more than one.

Optionally, the chip further includes a memory 1903 for storing necessary computer programs and data.

Those skilled in the art can also understand that various illustrative logical blocks and steps listed in the embodiments of the present application can be implemented by electronic hardware, computer software, or a combination of both. Whether such functions are implemented by hardware or software depends on the specific application and overall system design requirements. Those skilled in the art may use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the protection scope of the embodiments of the present application.

The embodiment of the present application also provides a system for compressing and feeding back channel state information CSI. The system includes the communication device as the terminal device (such as the first terminal device in the method embodiment above) and the network device as the communication device in the embodiment of FIG. 7 The communication device, alternatively, the system includes the communication device serving as the terminal device (such as the first terminal device in the foregoing method embodiment) and the communication device serving as the network device in the foregoing embodiment in FIG. 18 .

The present application also provides a readable storage medium on which instructions are stored, and when the instructions are executed by a computer, the functions of any one of the above method embodiments are realized.

The present application also provides a computer program product, which implements the functions of any one of the above method embodiments when executed by a computer.

In the above embodiments, all or part of them may be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer programs. When the computer program is loaded and executed on the computer, all or part of the processes or functions according to the embodiments of the present application will be generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer program can be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer program can be downloaded from a website, computer, server or data center Transmission to another website site, computer, server or data center by wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media. The available medium may be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a high-density digital video disc (digital video disc, DVD)), or a semiconductor medium (for example, a solid state disk (solid state disk, SSD)) etc.

Those of ordinary skill in the art can understand that: the first, second and other numbers involved in this application are only for the convenience of description, and are not used to limit the scope of the embodiments of this application, and also indicate the sequence.

At least one in this application can also be described as one or more, and multiple can be two, three, four or more, and this application does not make a limitation. In this embodiment of the application, for a technical feature, the technical feature is distinguished by "first", "second", "third", "A", "B", "C" and "D", etc. The technical features described in the "first", "second", "third", "A", "B", "C" and "D" have no sequence or order of magnitude among the technical features described.

The corresponding relationships shown in the tables in this application can be configured or predefined. The values of the information in each table are just examples, and may be configured as other values, which are not limited in this application. When configuring the corresponding relationship between the information and each parameter, it is not necessarily required to configure all the corresponding relationships shown in the tables. For example, in the table in this application, the corresponding relationship shown in some rows may not be configured. For another example, appropriate deformation adjustments can be made based on the above table, for example, splitting, merging, and so on. The names of the parameters shown in the titles of the above tables may also adopt other names understandable by the communication device, and the values or representations of the parameters may also be other values or representations understandable by the communication device. When the above tables are implemented, other data structures can also be used, for example, arrays, queues, containers, stacks, linear tables, pointers, linked lists, trees, graphs, structures, classes, heaps, hash tables or hash tables can be used wait.

Predefinition in this application can be understood as definition, predefinition, storage, prestorage, prenegotiation, preconfiguration, curing, or prefiring.

Those skilled in the art can appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present application.

Those skilled in the art can clearly understand that for the convenience and brevity of the description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

The above is only a specific implementation of the application, but the scope of protection of the application is not limited thereto. Anyone familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the application. Should be covered within the protection scope of this application. Therefore, the protection scope of the present application should be determined by the protection scope of the claims.

Claims

A method for channel state information CSI compression feedback, characterized in that it is applied to a terminal device, and the method includes:

Obtain an estimated CSI image H of the network device, and generate a time-series CSI image Hc according to the estimated CSI image H;

Compressing the time series CSI image Hc to generate a feature codeword;

Send the feature codeword to the network device.
The method according to claim 1, wherein the compressing the time-series CSI image Hc to obtain a feature codeword comprises:

The time-series CSI image Hc is input into the self-information domain converter to generate a time-series self-information image He , wherein the time-series CSI image Hc and the time-series self-information image He are both T in time dimension;

Input the time-series self-information image He into a time-series feature coupling encoder for feature extraction to generate a structural feature matrix and a temporal correlation matrix;

The feature codeword is generated according to the structural feature matrix and the time correlation matrix.
The method according to claim 2, wherein the input of the time-series CSI image Hc into a self-information domain converter to generate a time-series self-information image comprises:

Input the time-series CSI image Hc into the three-dimensional convolutional feature extraction network to extract features to obtain the first time-series feature image F, wherein the convolution kernel specification of the three-dimensional convolutional network is f×t×n×n, and the f is the number of feature extractions, the t is the depth of convolution in the time dimension, and the n is the length and width of the convolution window;

generating a first index matrix M according to the time series CSI image Hc ;

A time-series self-information image He is obtained according to the first time-series feature image F and the first index matrix M.
The method according to claim 3, wherein the generating the first index matrix M according to the time-series CSI image Hc comprises:

Input the time-series CSI image Hc into a self-information module to generate self-information of the area to be estimated in the time-series CSI image Hc , and use it as a self-information image;

The self-information image is input into the index matrix module for mapping to obtain the first index matrix M.
The method according to claim 4, wherein the step of inputting the time-series CSI image Hc into a self-information module to obtain self-information of the region to be estimated in the time-series CSI image Hc , to obtain a self-information image, includes :

Split the time-series CSI image H c in time series to obtain split images H c,i at each time point;

Divide the split image into multiple regions p j to be estimated, and obtain self-information estimation values of the regions to be estimated
According to the estimated value from the self-information
Generated from the information image I c,i .
The method according to claim 4, wherein the index matrix module includes a mapping modulus network and a decision device, and the mapping of the self-information image input index matrix module to obtain the first index matrix M includes:

Inputting the self-information image into the mapping network to extract features to obtain a first information feature image D c,i , wherein the mapping network is a two-dimensional convolutional neural network;

Inputting the first information feature image D c,i into the decision device for binarization processing to obtain a second index matrix M i ;

The second index matrix M i is concatenated to obtain the first index matrix M.
The method according to claim 6, wherein the mapping network includes a two-dimensional convolutional layer, a two-dimensional normalization layer and an activation layer, and the self-information image is input into the mapping network to extract features, include:

inputting the self-information image into the two-dimensional convolutional layer to extract features to obtain a first feature image;

inputting the first feature image into the two-dimensional normalization layer to normalize the pixel values in the first feature image to obtain a second feature image;

Inputting the second feature image into an activation function layer for nonlinear mapping to obtain the first information feature image D c,i .
The method according to claim 6, wherein said splicing said second index matrix Mi to obtain a first index matrix M comprises:

The second index matrix M i is spliced in a time series order to obtain the first index matrix M.
The method according to claim 3, wherein the acquiring a time-series self-information image according to the first time-series feature image F and the first index matrix M further comprises:

multiplying the first time-series feature image F and the first index matrix M to obtain a second information feature image;

Inputting the second information feature image into a dimension restoration network to perform dimension restoration to generate the time-series self-information image He .
The method according to claim 2, wherein the time series feature coupled encoder comprises a one-dimensional space-time compression network and a coupled long short-term memory network (LSTM).
The method according to claim 10, characterized in that, said time-series self-information image He is input into a time-series feature coupling encoder to perform feature extraction to generate a structural feature matrix and a temporal correlation matrix, comprising:

The time sequence is input into the one-dimensional space-time compression network for one-dimensional space-time compression after dimension transformation from the information image He to obtain the structural feature matrix, wherein the convolution kernel specification of the one-dimensional space-time compression network is S× 2N c N t ×m, the 2N c N t is the length of the convolution window, the m is the width of the convolution window, S is the target dimension, and the dimension of the structural feature matrix is T×S.
The method according to claim 10, wherein the described time series self-information image He is input into a time series feature coupling encoder to perform feature extraction to generate a structural feature matrix and a temporal correlation matrix, further comprising:

After performing dimension transformation on the time series from the information image He , input coupling LSTM to extract features to obtain the temporal correlation matrix, wherein the dimension of the temporal correlation matrix is T×S;

The structural feature matrix and the temporal correlation feature matrix are coupled to generate the feature codeword.
The method according to any one of claims 1-12, further comprising:

Input the training time-series CSI image Hc into the self-information domain converter to obtain the training time-series self-information image He ;

Input the training time-series self-information image He into the time-series feature coupling encoder to obtain the training feature codeword.
The method according to claim 13, further comprising:

Send the training data to the network device, wherein the training data includes the training feature codeword, the time sequence length of the time-series self-information image He , the dimension of the training feature codeword and the training time-series CSI image Hc .
A method for channel state information CSI compression feedback, characterized in that it is applied to network equipment, and the method includes:

Receive the feature code word sent by the terminal device;

Restoring the feature codeword to obtain the restored time-series CSI image

Restore timing CSI images according to the
Get the estimated CSI image of the end device
The method according to claim 15, wherein said restoring said characteristic codeword comprises:

Inputting the feature codeword into a time-series feature coupling decoder to obtain the restored time-series CSI image
The method according to claim 16, wherein the time-series feature coupling decoder includes a decoupling module and a restored convolutional neural network, and the acquisition of the restored time-series CSI image
include:

Input the feature codeword into the decoupling module for decoupling, so as to obtain the restored time series self-information image

Restore the time series from the information image
Input the restored convolutional neural network for restoration to obtain the restored time-series CSI image
The method according to claim 17, wherein the decoupling module includes a one-dimensional space-time decompression network and a decoupling LSTM, and the decoupling is performed by inputting the feature codeword into the decoupling module, so as to obtain the restored timing self Informational images, including:

Inputting the feature codeword into the one-dimensional space-time decompression network for decompression to obtain the restored structure feature matrix;

Inputting the feature codeword into the decoupling LSTM for decoupling to obtain the restored time correlation matrix;

Acquiring the restored time-series self-information image according to the restored structural feature matrix and the restored time correlation matrix
The method according to claim 17, wherein the convolution kernel specification of the one-dimensional space-time decompression network is 2N c N t ×s ×m, and the T is the number of rows of the restored time correlation matrix , the 2N c N t is the number of columns of the restored time correlation matrix.
The method according to claim 18, characterized in that the restoration time series self-information image is obtained according to the restoration structure feature matrix and the restoration time correlation matrix
include:

Adding the restored structural feature matrix and the restored time correlation matrix point-to-point, and performing dimension transformation to obtain the restored time-series self-information image
The method according to claim 17, wherein the restored convolutional neural network comprises a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer , the sixth convolutional layer, the seventh convolutional layer, wherein the convolution kernel specifications of the first convolutional layer and the fourth convolutional layer are l 1 ×t×n×n, and the second convolutional layer and the convolution kernel specification of the fifth convolution layer is l 2 ×t×n×n, and the convolution kernel specification of the third convolution layer, the sixth convolution layer and the seventh convolution layer is 2×t× n×n, the t is the depth of the convolution in the time dimension, the l 1 , l 2 and 2 are the number of extracted features, and the n is the length and width of the convolution window.
The method according to claim 21, wherein said restoring the time sequence from the information image
Input the restored convolutional neural network for restoration to obtain the restored time-series CSI image
include:

Restore the time series from the information image
Input the first convolutional layer for convolution to obtain a first restored feature map, input the first restored feature map to the second convolutional layer to obtain a second restored feature map, and input the second restored feature map The third convolutional layer is used to obtain a third restored feature map, and the third restored feature map and the restored time-series self-information image
sum to obtain a fourth reduced feature map;

Inputting the fourth restored feature map into the fourth convolutional layer to obtain a fifth restored feature map, inputting the fifth restored feature map into the fifth convolutional layer to obtain a sixth restored feature map, and converting the The sixth restored feature map is input into the sixth convolutional layer to obtain a seventh restored feature map, and the fourth restored feature map and the seventh restored feature map are added to obtain an eighth restored feature map;

Inputting the eighth restored feature map into the seventh convolutional layer for normalization to obtain the restored time-series CSI image
The method according to any one of claims 15-22, further comprising:

Receiving the training data sent by the terminal device, the training data includes the training feature codeword, the time sequence length of the time-series self-information image He , the dimension of the training feature codeword and the training time-series CSI image;

Obtaining a restored time-series CSI image according to the training feature codeword;

Perform training according to the restored time-series CSI images and the training time-series CSI images.
The method according to claim 23, further comprising:

Determine the number of structural units in the decoupled LSTM according to the time series length of the time series self-information image He ;

The network parameters of the one-dimensional space-time decompression network are determined according to the dimensions of the training feature codewords.
The method according to claim 23, further comprising:

Carry out multiple rounds of training, the formulation of the learning rate in the training is expressed as:

Wherein, the γ is the current learning rate, the γ max is the maximum learning rate, the γ min is the minimum learning rate, the t is the current training round, and the T w is the number of gradual learning, so The T' is the number of overall training cycles.
The method according to claim 23, further comprising:

Obtain recommended network parameters of the decoupling module and restored convolutional neural network, and update the decoupling module and restored convolutional neural network according to the recommended network parameters.
A communication device, characterized by comprising:

An estimation module, configured to acquire an estimated CSI image H of the terminal device, and generate a time-series CSI image Hc according to the estimated CSI image H;

A compression module, configured to compress the time series CSI image Hc to generate a feature codeword;

A sending module, configured to send the feature codeword to a network device.
A communication device, characterized by comprising:

The receiving module is used to receive the characteristic code word sent by the terminal equipment;

A restore module, configured to restore the feature codewords to obtain restored time-series CSI images

A channel acquisition module, configured to restore time series CSI images according to the
Get the restored estimated CSI image
A communication device, characterized in that the device includes a processor and a memory, and a computer program is stored in the memory, and the processor executes the computer program stored in the memory, so that the device performs the The method described in any one of 1 to 14.
A communication device, characterized in that the device includes a processor and a memory, and a computer program is stored in the memory, and the processor executes the computer program stored in the memory, so that the device performs the The method described in any one of 15-26.
A communication device, characterized by comprising: a processor and an interface circuit;

The interface circuit is used to receive code instructions and transmit them to the processor;

The processor is configured to run the code instructions to execute the method according to any one of claims 1-14.
A communication device, characterized by comprising: a processor and an interface circuit;

The interface circuit is used to receive code instructions and transmit them to the processor;

The processor is configured to run the code instructions to execute the method according to any one of claims 15-26.
A computer-readable storage medium is used for storing instructions, and when the instructions are executed, the method according to any one of claims 1-14 is realized.
A computer-readable storage medium for storing instructions, which, when executed, cause the method according to any one of claims 15-26 to be implemented.